The Neuron Sunday Special header image showing what to build with Fable 5

In partnership with

Welcome, humans.

Weekend homework: CAIS and Scale say Claude Fable 5 now scores 16.1% on the Remote Labor Index, a benchmark that tests whether models can do real freelance-style computer work.

That is impressive. It is also the kind of number that makes people immediately ask the practical question: cool, what should I actually build with it?

Our answer: use Fable like a weekend contractor, not a chatbot. Give it a real goal, a clean brief, and a finish line. Asking Fable to summarize a PDF is like hiring a Michelin chef to microwave leftovers.

Here’s what happened in AI today:

😺 Fable 5 became your weekend build contractor
📰 Synthetic cells grew and divided in the lab
📰 StoryScope spotted AI fiction by plot shape
🍪 Vellum launched a memory-first personal assistant
🎓 Blind-spot checks make AI sessions safer

… and a whole lot more that you can read about here.

P.S: Want to reach 700,000 AI-hungry readers? Click here to advertise with us.

😺 What To Build With Fable 5 This Weekend

The most useful way to think about Claude Fable 5 is simple: stop giving it errands and start giving it jobs.

Fable is the model people keep talking about because it appears unusually strong at long-horizon work, aka tasks that require many steps, course corrections, and judgment instead of one answer. CAIS and Scale said Fable 5 hit 16.1% on the Remote Labor Index, roughly doubling the next-best model on real remote-work tasks.

Here's what happened:

Fable 5 became the obvious model to try for ambitious builds, codebase cleanup, app cloning, and agentic workflows.
Anthropic's access situation stayed messy, with Fable temporarily off subscription plans after July 7 while capacity gets sorted.
Builders started treating Fable like a scarce resource: plan cheaply, execute carefully, then save the expensive run for the hard part.

How to try it: Pick one project with a visible finish line. Good candidates from Chase AI's Fable walkthrough include: clone a paid app you want to customize, build an internal dashboard, audit your Claude Code setup, repair a messy repo, or turn a product requirements document into working software.

Peter Yang released his own five use-cases: first ask Fable to inspect your memory, files, projects, or docs and tell you which tasks are actually Fable-worthy. Then use it for jobs where judgment compounds, like business planning, shipping audits, big-feature specs, or cleaning up your personal OS.

So, to reiterate, do at least one of the following:

Clone one paid app you wish worked differently, then customize the workflow.
Audit one messy tool setup and ask for a safer, cheaper model-routing plan.
Build one personal dashboard that pulls together your calendar, inbox, docs, and tasks.
Refactor one painful codebase with tests, checkpoints, and rollback notes.
Turn one PRD into software, then make Fable compare the result against the requirements before calling it done.

But before you do, do the boring prep with a cheaper model. Ask it to write the requirements, list risks, define success criteria, and produce the exact handoff prompt. Fable should get the complete job packet, not your half-formed brainstorm.

The key pattern: prep the boring context with a cheaper model, give Fable a plan doc plus any useful APIs/MCPs, let Fable plan or audit, then hand execution to a cheaper model when possible. Many users are recommending to use lower effort, which means you’ll need to babysit it a bit more, but this will prevent runaway loops that can burn your limit fast. Try this:

I want to use Fable 5 only where it is worth the tokens.

Context I can provide:
[projects, memory, files, plan doc, metrics, codebase, skills, APIs/MCPs]

First, inspect the context and list the top five Fable-worthy tasks. Prioritize:
1. Work that needs a large corpus of context.
2. Planning or advice that affects the next 3-12 months.
3. Ship-readiness audits for real projects.
4. Plans detailed enough for a cheaper model to execute.
5. Refactors or cleanup of my personal OS, skills, or codebase.

For each task, give:
- Why Fable is worth using.
- What context you still need.
- Whether Fable should plan, audit, or execute.
- How to prevent token waste or loops.

Then recommend the single highest-leverage task to run now.

Why this matters: Most people waste frontier models by asking them to think harder about small tasks. Fable's advantage shows up when the task has branches, uncertainty, and enough room for judgment to matter.

The practical shift is model routing. Use cheaper models for research, planning, and drafts (just maybe not Sonnet 5). Use Fable when execution quality changes the outcome.

Our take: Fable's best weekend use is the project you keep postponing because it has too many steps. Give it one clear build, then make the model prove it can finish. The counterpoint is cost: if the brief is vague, Fable will turn your indecision into an expensive progress bar. Still might be useful, but you gotta max those tokens while you get them at a subscription discount!

FROM OUR PARTNERS

Scale AI support on AWS, see how July 9

Customer expectations keep rising. Support budgets don't. On July 9, Fin and AWS are hosting a live executive session on how leading enterprises close that gap: scaling AI-powered support while simplifying how they buy it.

You'll see how to resolve an average 76% of conversations with Fin on AWS enterprise-grade infrastructure, procure through AWS Marketplace to put committed cloud spend to work, and turn the Fin and AWS collaboration into lower support costs. Register for the live session to see how.

Save your spot

🎓 AI Skill of the Day: End Every AI Session With A Blind-Spot Check

Your best AI answers usually fail in the quiet parts: the assumptions it skipped, the risks it underweighted, and the thing you forgot to ask.

A useful ClaudeAI workflow thread recommends ending sessions with two audit questions. The trick is to make the model critique both itself and your framing before you act on the answer.

Use this after a strategy doc, code plan, vendor decision, research summary, or anything where being confidently incomplete would hurt.

Before we finish, run a blind-spot audit.

1. What part of your answer are you least confident about, and why?
2. What am I missing about this situation?
3. What assumption would most change your recommendation if it were wrong?
4. What should I verify with a human, source, log, or test before acting?

Be specific. Do not reassure me. Give me the risk, the evidence gap, and the next check.

Total AI beginner? Start here (goes with this video).

Have a specific skill you want to learn? Request it here.

🍪 Treats to Try

Vellum gives you a personal assistant with evolving memory, task handling, and preferences, so it can coordinate work in Slack like a teammate, with no pricing details listed.
ZCode gives developers a free coding agent built on Z.ai's GLM model, with repo editing, terminal commands, and Cursor/Claude Code-style workflows, with no pricing details listed.
Seedance 2.5 in Dreamina turns prompts and up to 50 multimodal references into 30-second cinematic videos with R2V control, with no pricing details listed.
Context.dev converts websites into simple markdown, HTML, or structured data for agents, with JavaScript rendering and site-wide crawling, with no pricing details listed.
Safari MCP server connects agents to a real Safari Technology Preview window to inspect pages, capture screenshots, and debug web apps, with no pricing details listed.

📰 Around the Horn

Axios reported that Anthropic's Fable/Mythos revival involved a 20-day scramble across Amazon, Commerce, CAISI, NSA, Treasury, and the White House as OpenAI kept negotiating GPT-5.6 release terms.
Quanta reported that researchers built synthetic cells with lab-made DNA that can feed, grow, copy genetic material, and divide.
StoryScope found AI-written fiction can be detected by plot structure, not just word choice, because AI stories over-explain themes and follow narrower arcs.
WorldModelGym introduced a benchmark that tests whether world models pick actions that actually win in the real environment.
SWE-Together released an interactive coding-agent benchmark based on real multi-turn coding sessions, with 109 tasks and a public leaderboard.
EdgeBench added long-horizon executable agent tasks that measure how agents improve over 12 to 72 hours.
Data centers meet heat waves as extreme heat adds strain to host communities, power grids, and the debate over who pays for AI's compute buildout.
H1 venture funding hit a record $510B globally, with AI money still concentrating around frontier labs, infrastructure, defense, robotics, and healthcare.

FROM OUR PARTNERS

Usage-based pricing is reshaping B2B revenue.Usage-based pricing is reshaping B2B revenue. Watch Tabs + PwC break down the finance ops reality — and how to scale it without adding headcount.

Watch it now

🌟 Sunday Special: The Week In AI

The short version: AI joins hardware teams, pricing models, government stakes, enterprise deployment, and tools that actually sit inside work.

Top 5 stories of the week:

OpenAI poached Apple's Vision Pro hardware lead, adding another ex-Apple operator to its Jony Ive-led device push while Apple keeps trying to make its own AI hardware feel like Apple.
AI put more pressure on billable hours, as consulting clients pushed firms toward fixed-fee and outcome-based pricing now that AI can compress work that used to be sold by the hour.
Anthropic launched Claude Sonnet 5, making its stronger everyday agent model the default for Free and Pro users while also launching Claude Science for research workflows.
Fable 5 came back online after a government-triggered shutdown, then immediately became the week's model-routing obsession: use it for planning, judgment, and hard reviews, not every tiny task.
OpenAI reportedly discussed giving the U.S. government a 5% stake, turning the AI upside debate from abstract policy chatter into a very real ownership question.

Top 5 tools worth trying:

Claude Fable 5 is the scarce, expensive model to use for planning, judgment, code reviews, and hard project audits while the access window is still open.
Claude Science gives researchers a beta workbench with code-traced artifacts, on-demand environments, and 60+ optional scientific database connectors.
Cursor for iOS lets you launch cloud coding agents, control desktop agents, review diffs, and merge PRs from your phone.
Google's Nano Banana 2 Lite and Gemini Omni Flash pushed cheaper image generation and developer video editing further into the Gemini stack.
The open coding stack got deeper: Qwen-AgentWorld gives developers an open-source agent training and evaluation environment, ZCode brings a free GLM-powered coding agent, Kimi 2.7 Code adds autonomous goal execution, and ClinePass gives cheaper access to open coding models inside Cline.